What concurrent request threshold triggers LinkedIn's internal gateway to queue requests and cause latency spikes?

Answer to the Core Question

LinkedIn does not publish a fixed number for the concurrent‑request threshold that forces its internal gateway to queue traffic. Observations from similar API gateways suggest the switch to queuing occurs somewhere between a few hundred and several thousand simultaneous connections per gateway instance. The exact value is dynamic and can shift with traffic patterns, geographic region, and the specific API product in use.

Confirmed Facts

• LinkedIn’s public docs only expose per‑application rate limits (e.g., 100 req/s). The gateway’s queuing behavior is an internal, adaptive control mechanism.
• Latency spikes attributed to queuing appear when the 99th‑percentile response time jumps sharply while HTTP 429 responses remain unchanged.
• The threshold can differ across core REST APIs, Marketing Developer Platform, and Learning APIs, and may vary by region (us‑west vs. eu‑central).

Likely Explanation

Large‑scale API gateways typically monitor the number of in‑flight requests and, when that number exceeds a configurable limit, they place excess traffic in a short‑term queue. The queue length is usually bounded by the gateway’s ability to keep request ordering and fairness. Because LinkedIn’s infrastructure is highly elastic, the exact limit is not static; it can be adjusted on the fly based on real‑time load, upstream service health, and regional traffic distribution.

Practical Steps to Pinpoint the Threshold

1. Choose the API endpoint and region you want to test (e.g., v2/ugcPosts in us‑west).

2. Set up a load‑generation script that ramps concurrent requests in increments (e.g., 50, 100, 200, 400, 800, 1600).

3. For each increment, record the following metrics:

   • 99th‑percentile latency
   • Average latency
   • HTTP 429 (Too Many Requests) count
   • Any LinkedIn‑specific headers indicating queue status (e.g., X-LinkedIn-Queue-Depth if present)

4. Identify the concurrency level where the 99th‑percentile latency shows a sudden, non‑linear increase while the 429 count remains flat. That point is a strong indicator of gateway queuing.

5. Repeat the test for other endpoints and regions to observe variability.

Diagnostic Detail Needed for Tailored Guidance

To refine the recommendation, I need to know which specific API endpoint and geographic region you are targeting. The queuing threshold can differ significantly between, for example, v2/adCampaignsV2 in eu‑central and v2/learning in ap‑south‑east.